It was a debacle for pollsters, when most opinion polls failed to predict a landslide victory in the 2015 Bihar assembly elections for the mahagathbandhan. A majority of them even predicted the wrong winner. The average of 13 opinion poll results showed 113.5 (out of 243) seats for the Janata Dal (United), Rashtriya Janata Dal and Congress combined — a whopping 26.5% error, as these parties bagged 178 seats in total. While opinion polls go horribly wrong in numerous elections worldwide, this was a horrendous performance by any standard.
The number of available opinion polls for Bihar maybe less this time around; the 2015 blunder may be one reason for such temperance. To keep the error of any survey-based statistical estimation within reasonable limits — and to estimate the quantum of possible error — one must design, implement and carry out the subsequent inference with due care. Public disclosures of sampling method, sample profile and method used to convert votes into seats being typically unavailable in most election surveys, do most pollsters in India conduct their polls following appropriate statistical theory?
Usually, a ‘3% margin of error’ is targeted in an election survey. This means that there’s a 95% chance that the poll survey result will be within 3% of the actual election result. It can be shown that, in a homogeneous set-up, only 1,004 samples can achieve this, irrespective of the population size. However, the samples should be ‘random’, representing the population by approximately maintaining the population proportions across gender, age, income, religion, caste, etc.
India’s social structure, politics and multi-party democracy, however, make this task daunting. Constituencies being widely heterogeneous, ideally, one needs 1,004 samples in each constituency. Some similar-behaving constituencies may, however, be clubbed together to frame different homogeneous populations to reduce the sample size.
The above calculation corresponds to vote shares only. The situation becomes much more complicated in a first-past-the-post (FPTP) system like ours, where the number of seats has a very complicated non-linear relationship with vote shares, which is much more difficult to estimate. The 2016 US presidential election is a classic example how a crashing defeat in the electoral college may happen despite a 2.1% lead in popular votes. Hillary Clinton can vouch for that.
The number of seats is a very sensitive calculation. A small shift in vote shares may drastically change the dynamics in an FPTP system. In a two-party set-up, if both parties are poised at 50% vote share, the number of seats for the parties is expected to be same. With a 1% shift in vote share — 51% and 49% vote shares for the two parties — the winner would get 57.9% seats on an average. And a 5% shift in vote share in either direction would give 84.3% seats to the winner, on an average.
The situation would become much more complicated in a multi-party democracy like India’s with so many alliances, break-ups in alliances, not to mention leaders. Unless these issues are properly addressed in the poll prediction model, the overall performance of pollsters will once again look grim.
Poll predictions during Covid-19 have become even more tricky. Overwhelmed healthcare systems, shrinking economies and job losses have made any election anywhere a referendum on Covid management of the incumbent, at least partially. The landslide victory of Prime Minister Jacinda Ardern in New Zealand earlier this month is testament to that.
Also, election campaigning has, to a great extent, become virtual these days, providing a serious handicap for pollsters to connect candidates and voters. The pandemic has invariably triggered the increased use of social media, its notoriety in spreading misinformation being well-known.
Voter turnout might also be seriously affected during this pandemic. While election turnouts in Poland, Sri Lanka and New Zealand increased from the previous elections, in Croatia and Serbia they were lower than earlier occasions, with the turnout in Bolivia remaining almost constant. In the first phase of the Bihar election on Wednesday, the turnout was 54.3%, only a small dip from the 54.9% turnout in the first phase of polling in 2015.
A significant change in voter turnout due to Covid, though, could significantly influence the election, unless this occurs proportionately across all socioeconomic and demographic groups. An increase in voter turnout in the US elections, for instance, could be due to additional enthusiasm of younger people, ‘non-White’ and lower-income groups — mostly Democrat voters. Again, as Americans above 65 support Republicans more, if fewer senior citizens come out to vote on November 3 due to Covid-19, this could hamper the prospect of President Donald Trump.
Have our pollsters taken similar issues into consideration for India’s first Covid-era election in Bihar? Even if some opinion polls will throw up reasonably accurate predictions, I’d prefer putting up a statutory warning regarding these polls: for entertainment only.
The writer is professor of statistics, Indian Statistical Institute, Kolkata